Robots are getting better at reasoning and planning, yet they still stumble on a more basic question: where, exactly, is the thing I am supposed to pick up? On November 4, 2026, XPENG announced GroundingPI, an open-source visual perception model built with researchers from Peking University, The University of Hong Kong and other institutions, that tackles that question head-on.
According to XPENG’s announcement, the roughly 4-billion-parameter model reached an average score of 73.68 across 34 perception benchmarks, ahead of the 71.54 the team recorded for GPT-6 Astra in the same evaluation. The paper, code and model weights are all public.
What GroundingPI is
GroundingPI is a visual foundation model focused on one fundamental skill: identifying objects and pinpointing their positions. XPENG describes it as a perception-focused approach to Physical AI, meaning AI that acts in the real world through robots and vehicles rather than only in text.
The press release gives a simple example. If an instruction asks for “the handle of the cup on the right,” a system must find the correct cup and then the precise location of its handle. GroundingPI represents such answers as spatial coordinates, including points and bounding boxes.
It was trained across 11 categories of perceptual capability, including object detection, dense object localization, referring expression grounding, optical character recognition, document layout understanding, user interface element recognition and robotic spatial localization. Notably, the perception model was trained without robot action labels. Its job is to supply a strong visual foundation that action-learning systems can build on later.
The benchmark results, and their limits
The team compared GroundingPI with 44 baseline models on 34 public benchmarks covering those 11 categories. Using an aggregate score that weights each category equally, GroundingPI reached 73.68 against 71.54 for GPT-6 Astra. The model did particularly well on tasks that demand precise localization, such as busy scenes with many objects or very small targets.
XPENG is careful about what this does and does not show. The company states that the comparison relates only to the evaluated perception benchmarks and does not represent a broader assessment of general intelligence or reasoning. In other words, a specialist beat a larger generalist at a specialist’s job, which is the point of the research rather than a claim of overall superiority.
Fewer demonstrations, better generalization
The more interesting results concern robot learning. In RoboCasa-GR1 simulation experiments, systems that used GroundingPI as their visual backbone achieved higher task success with 50 percent of the available demonstration data than the comparison models trained on 75 percent. Collecting human demonstrations is one of the costliest parts of teaching robots new manipulation tasks, so a stronger perceptual starting point could reduce that burden.
In RoboTwin 2.0 simulation, GroundingPI outperformed the evaluated mainstream visual backbones across four out-of-distribution settings, where objects, backgrounds or conditions differ from training. The team also plugged it into an autonomous driving setup on the nuScenes dataset and reported an average open-loop trajectory prediction error of 0.296 meters.
The release is explicit that the work is exploratory. The downstream results come from simulation and offline evaluations and do not establish readiness for real-world commercial deployment, nor do they announce integration into XPENG products.
What it means
For robotics developers, the practical appeal is the open release. Anyone can download the weights and test whether a perception-first backbone helps their own manipulation or navigation stack, subject to the open-source and third-party licenses noted by the team.
For the wider AI field, GroundingPI is a data point for a modular view of embodied intelligence: pair a sharp, specialized perception layer with general-purpose reasoning and planning rather than asking one giant model to do everything. Whether that holds up on real hardware is the open question, and independent replication will matter.
Bottom line
GroundingPI will not put a smarter robot in your kitchen next month. What it offers is a well-documented, openly shared step toward machines that see precisely, and early evidence that doing so could cut the training data robots need. Watch for real-world tests and outside benchmarks as the next signals.